Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts

Yu Cao^* ,  Zengqun Zhao ,  Ioannis Patras ,  Shaogang Gong^†

Queen Mary University of London

ASCED provides method for detecting and correcting artifacts in diffusion-generated images through temporal score analysis. The repository includes two main demonstration notebooks:

• notebooks/detection_demo.ipynb: Demonstrates artifact detection in generated images
• notebooks/correction_demo.ipynb: Demonstrates artifact correction in diffusion models

Before running the notebooks, you need to download:

1. Model weights: Download the pre-trained diffusion model weights from yandex-research/ddpm-segmentation and place them in the checkpoints/ddpm/ directory
2. Pickle files and seed data: Download the following from HERE:
   • normalized_score_dict.pkl and place it in experiments/
   • Seed files (noise_*.npy) and place them in datasets/noise/

1. Clone the repository:

git clone https://github.com/YuCao16/ASCED.git
cd ASCED

2. Install the required dependencies:

pip install -r requirements.txt

Open and run notebooks/detection_demo.ipynb to see demonstrations of:

• Temporal difference analysis
• Artifact mask generation
• Acceleration comparison between artifact and non-artifact regions

Open and run notebooks/correction_demo.ipynb to see demonstrations of:

• DDIM sampling with artifact correction
• Visual comparison of corrected outputs

If you find this work useful, please cite:

@inproceedings{cao2025temporal,
  title={Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts},
  author={Cao, Yu and Zhao, Zengqun and Patras, Ioannis and Gong, Shaogang},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={7707--7716},
  year={2025}
}

Please feel free to open an issue on GitHub if you encounter problems or have suggestions.

This project is licensed under the MIT License - see the LICENSE file for details.

Parts of this project page were adopted from the Nerfies page.